1 / 48

Constraint Models for the Covering Test Problems

Constraint Models for the Covering Test Problems. Authors: Brahim Hnich , Steven D. Prestwich , Evgeny Selensky , Barbara M. Smith Speaker: Pingyu Zhang CSE 990 Advanced Constraint Processing, Fall 2009. Outline. Introduction & background CSP model

leon
Télécharger la présentation

Constraint Models for the Covering Test Problems

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript

  1. Constraint Models for the Covering Test Problems Authors: BrahimHnich, Steven D. Prestwich, EvgenySelensky, Barbara M. Smith Speaker: Pingyu Zhang CSE 990 Advanced Constraint Processing, Fall 2009

  2. Outline • Introduction & background • CSP model • Naïve model, Alternative model, Integrated model • Symmetry breaking • Solving the CSP • Setting up the Ilog Solver, Experiment results • SAT Model • Model, Experiment results

  3. The Combinatorial Testing Problem • A combinatorial problem • A machine with 10 switches, each with two positions • 210 possible combinations to be tested before shipping! • Real world problems • Software interaction testing • Hardware testing • Testing of advanced materials • Interactions regulating gene expressions

  4. Optimization Problem: CA(t,k,g) • Given • A set of k software ‘components’ to be tested • Each component can take g different values (alphabet) • A covering strength t • Find • the minimum number of tests, b • Such that every combination of t components appears in some test k=5 b=2 b=? t=3

  5. Mathematical Object: Covering Array • Given: CA(t, k, g) of size b: b×k array • k: # of parameters • g: size of the alphabet • t: covering strength: for every subset of t parameters, each value combination appear at least once • Example CA(3, 5, 2), 25=32? b=10 does it! • Goal: Minimize b – Unfortunately, NP hard

  6. How about solving it with CSP • Challenges • How to model the problem into a CSP? • How to setup a solver? • The general framework • Step 1: guess a CA size b • Step 2: encode the problem as a CSP and try to solve it • Step 3: • if the CSP is solvable, reduce b by some value and repeat Steps 1—3 . • If CSP is not solvable, enlarge b by some value and repeat Steps 1—3. • The smallest solvable b is the optimum size.

  7. Outline • Introduction & Background • CSP model • Naïve model, Alternative model, Integrated model • Symmetry breaking • Solving the CSP • Setting up the Ilog Solver, Experiment results • SAT Model • Model, Experiment results

  8. A Naïve Matrix Model • Rows Each variable represents the value of one parameter in one vector: b×k matrix • Example: k=5 (xi1..5), g=2 (0/1), t=3 (every combination of three parameters), b=5 (a total of 5 combinations/rows) • How to express constraints? • Every subset of t parameters must be combined in all possible ways

  9. A Naïve Matrix Model • Each variable represent the value of one parameter in one vector: b×k matrix • How to express constraints? • Every subset of t parameters must be combined in all possible ways • Introduce additional boolean variables Describes the situation where x1, x2, x3 are assigned 0 at vector 1

  10. A Naïve Matrix Model • Each variable represent the value of one parameter in one vector: b×k matrix • How to express constraints? • Every subset of t parameters must be combined in all possible ways

  11. A Naïve Matrix Model • Each variable represent the value of one parameter in one vector: b×k matrix • How to express constraints? • Every subset of t parameters must be combined in all possible ways • Question: How many columns are there for this example?

  12. A Naïve Matrix Model • Each variable represent the value of one parameter in one vector: b×k matrix • How to express constraints? • Every subset of t parameters must be combined in all possible ways. • Express coverage constraints over columns (in general, columns) • Clearly, this is not very efficient…

  13. Alternative Matrix Model • What did we learn from last model? • Setting up variables is simple and straightforward. • Expressing constraints is cumbersome and introduces numerous assistant variables. • Each variable represents a t-tuple in a vector (compound variable), domain is all possible values for the tuple.

  14. Alternative Matrix Model • How to express constraints? • Now, each column represents all assignments for one specific tuple. • In each column, every number in the range 0 to 2t -1 should appear at least once, and at most… • Question: How many times each number appears at most in a column? • Enforcing such a global constraint on each column. ( constraints in total)

  15. Alternative Matrix Model • Improvement over the last model? • Are we done?

  16. Alternative Matrix Model • Intersection Constraints • Values assigned to compound variables in the same vector should be consistent. • In every row, if two variables share some positions, we impose a binary constraint stating the allowed pairs of values.

  17. Alternative Matrix Model • How many Intersection Constraints for each row? • For each compound variable, # of variables sharing ONE position is • # of variables sharing TWO positions • Continues until the case of sharing (t-1) positions • For the CA(3, 5, 2) example • # of variables sharing ONE position is 3 • # of variables sharing TWO positions is 6 • 10 variables in each row, (3+6)×10/2 = 45 constraints.

  18. Alternative Matrix Model • How many Intersection Constraints for each row? • Reduction 1: the red constraint is redundant with the conjunction of the two blue constraints and it can be removed. • In general, constraints between variables that share fewer than t-1 positions in common, can be removed. 1,4 1,2 1

  19. Alternative Matrix Model • How many Intersection Constraints for each row? • Reduction 2: a clique can be reduced to a path. • The savings are more obvious if k gets bigger. 1,2 1,2 1,2 (1, 2, 3) (1, 2, 3) (1, 2, 4) (1, 2, 4) (1, 2, 7) (1, 2, 7) (1, 2, 6) (1, 2, 5) (1, 2, 6) (1, 2, 5)

  20. Alternative Matrix Model • How many Intersection Constraints for each row? • Apply reduction 1: 45  30

  21. Alternative Matrix Model • How many Intersection Constraints for each row? • Apply reduction 1: 45  30 • Apply reduction 2: 30  20

  22. Alternative Matrix Model • Still, the # of intersection constraints is a problem  • Why not apply the intersection constraints on the naïve model? • To summarize so far • Naïve model suffers from # of additional variables. • The alternative model suffers from # of intersection constraints. • Can we combine the strengths together? Implicit connections

  23. Combining them together: an Integrated Model • Channeling constraint • explicitly express connections between compound variable and its corresponding t variables in the naïve model. • There are totally channeling constraints, each with arity t+1. • Enforcing GAC on them leads to assignment of x or domain reduce of c • Eliminate all intersection constraints.

  24. Outline • Introduction & Background • CSP model • Naïve model, Alternative model, Integrated model • Symmetry breaking • Solving the CSP • Setting up the Ilog Solver, Experiment results • SAT Model • Model, Experiment results

  25. Symmetry • CA has both row and column symmetries: any two cols/rows • Another less obvious symmetry: value symmetry • Size of the symmetry group: • So we must deal with them!

  26. Symmetry Breaking • For row and column symmetries • [Flener02] Theorem1: For a matrix model with row and column symmetry in some 2-d matrix, each symmetry class of assignments has an element where both the rows and the columns of that matrix are lexicographically ordered. • Naïve model: Impose lexicographical ordering constraints on both rows and columns of the matrix • Alternative model: rows are the same, columns are not as powerful, but can be help • Integrated model: can apply lexicographical ordering on either original or alternative matrix, but not both.

  27. Symmetry Breaking • For value symmetries • Very simple: force the first row to be all 0s (no matter which model) • It does not conflict with row/column symmetry-breaking constraints (the first row is lexicographically smallest anyway) • Triggers better propagation.

  28. Outline • Introduction & Background • CSP model • Naïve model, Alternative model, Integrated model • Symmetry breaking • Solving the CSP • Setting up the Ilog Solver, Experiment results • SAT Model • Model, Experiment results

  29. Setting up the solver • Use ILOG Solver 6.0 • Use compound variables as search variables • Coverage constraints expressed over compound variables • Variable and value ordering • Values assigned in ascending order • Labeling variables by column (coverage constraints are on columns) • Lexicographical ordering within each column • Symmetry-breaking constraints (as shown before) • Channeling constraints • Extensional: GAC • Intensional: bounds consistency • Same for domain of binary alphabet: each value is a bound. However, solver may run faster on bounds consistency

  30. Experiment results • Results for CA(3, k, 2) • For alternative model, both with all/minimal intersection constraints • For integrated model, channeling constraints expressed both extensionally (GAC) and intensionally (bounds consistency). • * indicates finding optimal size of b.

  31. Experiment results • Results for CA(3, k, 2) • For integrated model, enforcing intensional constraints is cheaper than enforcing extensional constraints.

  32. Experiment results • Results for CA(3, k, 2) • Integrated model is better than alternative model • For k=12, integrated model has 220 channeling constraints, while alternative model has 594 non-redundant intersection constraints.

  33. Experiment results • Results for CA(4, k, 2) • Runtime limit is 1 hour (CA(4, 7, 2) for b=24 can’t finish in 1 hour).

  34. Outline • Introduction & Background • CSP model • Naïve model, Alternative model, Integrated model • Symmetry breaking • Solving the CSP • Setting up the Ilog Solver, Experiment results • SAT Model • Model, Experiment results

  35. Local search model with SAT • Encoding the problem into CNF SAT (Example CA(2, 3, 2)) • Still have two matrices: original and alternative • Create a Boolean variable for EACH value of EACH variable. (1,2), (2,3), (1,3)

  36. Local search model with SAT (1) • Encoding the problem into CNF SAT (2) (1,2), (2,3), (1,3)

  37. Local search model with SAT (1) • Encoding the problem into CNF SAT (2) (3) (4) (1,2), (2,3), (1,3)

  38. Local search model with SAT (1) • Encoding the problem into CNF SAT (2) (3) (4) (5) (1,2), (2,3), (1,3)

  39. Local search model with SAT (1) • Encoding the problem into CNF SAT (2) (3) (4) (6) (5) (1,2), (2,3), (1,3)

  40. Local search model with SAT (1) • Encoding the problem into CNF SAT (2) (3) (4) (6) (5) (1,2), (2,3), (1,3)

  41. Local search model with SAT (1) • Encoding the problem into CNF SAT (2) (3) (4) (6) (5) (1,2), (2,3), (1,3)

  42. Setting up the solver • SAT local search algorithm • First, randomly assign all boolean variables. • Select a violated clause, flip the value of one variable in it (satisfy the clause, but may cause others to violate). • Repeat until timeout or satisfied. • Variable selection heuristic • With probability p, randomly select a variable. (noise, set to .2, .3, .4) • With 1-p, pick variable with min score • Break ties by selecting the variable that was least recently flipped. • Freebie move: if there are multiple variables such that flipping them would not create new unsat clauses (score=0), break ties randomly. Smallest negative effect Prefer variables that have fewest flips # of unsatisfied clauses results from flipping v Score = total # of flips in the past on all variables other than v

  43. Results for local search model: CA(2, k, g) *bold denotes best results. HW: HWFSAT (2006) HR: Hartmann & Raskin 2004 CK: Chateaunef & Kreher 2002 MS: Meagher & Stevens 2005 NU: Nurmela 2004

  44. Results for local search model • CA(3, k, g) • CA(4, k, g) Better than the other 2 In fact, they are optimal, but do we know they are optimal?

  45. Summary • Modeling CA problem into CSP • Naïve model • Alternative model • Integrated model • Local search with SAT • Handles larger problems • Trade scalability with completeness • Combining results with CSP results may be valuable

  46. Small Typos in Paper • Page 203, first formula: • Page 203, fifth line from bottom: variables • Page 212, 20th line from top:

  47. Literature • References • Hnich et al. ‘06: Constraints models for the covering test problems • Flener et al. ‘02: Breaking row and column symmetry in matrix models • Frisch et al. ‘02: Global constraints for lexicographical orderings • Gent et al. ‘95: Unsatisfied variables in local search • Hartman et al. ‘04:Problems and algorithms for covering arrays • Chateauneuf et al. ‘02: On the state of strength-three covering arrays • Meagher et al. ‘05: Group construction of covering arrays • Nurmela et al. ‘04: Upper bounds for covering arrays by Tabu search • The switches and light bulbs shown on first slide is a real game • http://www.devicedaily.com/gadgets/awesome-wooden-switch-and-bulb-game-resurrects-1960s-idea.html

  48. Discussion

More Related